What happened
According to 36Kr, US enterprise AI procurement has fractured into three distinct tiers driven by cost pressures. Large tech companies like Microsoft and Meta are restricting internal use of external models to favor in-house alternatives, mid-sized firms like Uber are slashing budgets due to rapid cost accumulation, and small startups are shifting to Chinese open-source models like DeepSeek and Tongyi Qianwen, which are significantly cheaper.
36Kr reports that the initial 'TokenMaxxing' trend, where companies encouraged high AI usage, has reversed as bills have become unsustainable. Uber, for example, exhausted its entire 2026 AI programming budget in under four months after deploying Claude Code to 5,000 engineers, with individual monthly costs reaching $500 to $2,000.
In response, the first tier of giants is building internal alternatives. Microsoft has reportedly cut its internal Anthropic budget by more than a third, shifting engineers to its own MAI-Code models. Meta has restricted employee use of Claude Code and OpenAI's Codex, reducing internal Claude users from 60,000 to 30,000, partly to avoid model issues and control costs.
The second tier, consisting of mid-sized enterprises, is facing budget cuts and service downgrades. Uber's CTO admitted the budget burn was unexpected, leading to strict hierarchical management of AI usage. Gartner predicts 25% of 2026 AI budgets will be deferred to 2027, largely driven by this tier's inability to justify the return on investment.
The third tier, comprising startups and small firms, is turning to Chinese open-source models. On the OpenRouter platform, Chinese models have reached a peak weekly share of 46.4%, with DeepSeek leading at 17.6%. These models are 60% to 90% cheaper than US flagship products, with specific price differences ranging from 4x to 100x, making them the default choice for cost-sensitive developers.
In reaction to this market shift, OpenAI has implemented two rounds of significant price cuts. In July, GPT-5.6 Luna was reduced by 80%, and in September, GPT-6 Sol was priced at half the cost of Anthropic's competing model. OpenAI attributes these cuts to internal efficiency improvements, including hardware routing and software optimizations.
Why it matters
This stratification indicates a fundamental shift in the AI market dynamics where unit price decreases are outpaced by increased consumption per task. US AI labs are losing market share at both the high end (to internal corporate models) and the low end (to Chinese open-source competitors). This forces US providers to engage in aggressive price cuts and emphasizes 'intelligence per dollar' over raw capability, potentially reshaping the competitive landscape and enterprise adoption strategies for the next decade.
The core issue is that while unit prices are falling, the volume of tokens consumed per task is rising faster, leading to higher overall bills. This 'cheaper token, more expensive bill' paradox is forcing a re-evaluation of AI ROI across all enterprise sizes.
US AI labs are experiencing a squeeze: they are losing their largest corporate clients to in-house development and their smallest clients to cheaper Chinese open-source alternatives. This dual loss threatens the long-term customer base of US labs, as startups adopting Chinese models now may not switch back as they grow.
The shift also highlights a compliance gap. Mid-sized US companies are hesitant to adopt Chinese models due to data security and regulatory uncertainties, even when the price advantage is substantial. This creates a market space where US labs can still compete, but only if they address cost concerns effectively.
The competitive dynamic is now centered on 'intelligence per dollar' rather than absolute capability. US labs are no longer competing on raw performance alone but must justify their premium pricing through efficiency and integrated value, a significant strategic pivot from previous years.
Interactive Mechanism: How It Actually Works
Explore the underlying technology behind this development interactively.
Which component of an AI application is the machine-learning model itself?
What to watch next
Monitor whether mid-sized US enterprises begin adopting Chinese open-source models despite compliance concerns, and track if US labs' price cuts successfully retain startup customers who might otherwise default to cheaper alternatives. Also watch for regulatory clarity on Chinese open-source AI usage in the US.
Watch for regulatory developments regarding the use of Chinese open-source AI models in US enterprises. Current ambiguity is keeping mid-sized firms from switching, but clearer rules could accelerate adoption.
Monitor the retention rates of startup customers for US AI labs. If OpenAI and Anthropic fail to retain these users despite price cuts, they risk losing the next generation of enterprise clients.
Observe whether the internal model strategies of giants like Microsoft and Meta become successful enough to reduce their reliance on external AI providers, potentially shrinking the market for US labs further.
Track the actual ROI of AI deployments. Gartner's prediction that only 28% of AI infrastructure projects have delivered their original business case suggests that many companies may continue to cut budgets if value is not demonstrated.